User Testing

Every automated test passes. Then Asha watches her neighbour use Pantry Pal for the first time. He types "some veg, leftover dal", stares at three suggestions, and asks, "Wait, does it know I'm vegetarian?" None of her tests could have told her that.

The rule

Give a real task, stay silent, and write down every stall.

User testing means watching real people use the working product, to find where it confuses them. It is different from the earlier validation: validation asked "is this worth building?", user testing asks "can people use what we built?"

How to run a session:

• Recruit from the target group: people who cook after work, not your developer friends. Around five people per round is usually enough to surface the biggest problems.
• Give a task, not instructions. "You have eggs, bread and spinach. Find something for dinner." Never "tap the green button".
• Ask them to think aloud, then stay quiet. When they get stuck, ask "What are you looking for?" instead of rescuing them.
• Write down every stall: a pause, a wrong tap, a misread word, a question. Record what happened, not your reading of it: "paused eight seconds at the diet question" is a finding, while "seemed confused" is not.

For an AI product, watch trust closely. Do people cook a suggestion without checking it? Do they ignore good ones? Both mean the design is not showing clearly what the AI knows and what it does not.

Then fix the biggest problem and run another round with new people. Each round is one more turn of the creation cycle.

The shortcut

Stall → cause → smallest fix → retest. For each stall, ask why it happened, choose the smallest change that removes it, and watch whether the next group stalls there too.

Drills

Answer fast — trust the rule.

Remember

• Give a real task, stay silent, and write down every stall.
• Stall → cause → smallest fix → retest with new people.
• Trap: rescuing testers, testing only with friends, and ignoring how much people trust the AI.